278

Papers

10/10

Roadblocks Active

0

Connections

Researchers built nine clinically-validated user personas representing depression, anxiety, PTSD, eating disorders, and incel identity, then deployed them against the Replika AI companion app. The app frequently mirrored or normalized unsafe content including self-harm and violent fantasy narratives, and showed a narrow emotional range dominated by curiosity and care regardless of user risk state. This matters because millions of vulnerable users interact with such tools under the assumption they provide a safe space, but the system appears to lack the clinical judgment to avoid reinforcing dangerous ideation.

██████████ 0.9 digital-therapeutics Preprint

ReadSaveConnections

This paper extracts features from how vocal state trajectories revisit similar patterns over time—rather than averaging acoustic levels—and shows these recurrence-based features outperform static acoustic baselines for detecting depression from conversational speech, achieving a mean AUC of 0.689. The key insight is that depression may alter the dynamic structure of how speech evolves, not just its average pitch or energy. Reproducibility is limited since no code or feature files are shared, but the approach offers a mechanistically motivated alternative to standard acoustic feature extraction.

██████████ 0.9 depression-biomarkers Preprint

ReadSaveConnections

A companion study to the recurrence paper above, this work finds that Shannon entropy of utterance-level acoustic trajectories (AUC 0.646) substantially outperforms static acoustic pooling (AUC 0.593) for depression classification on the same DAIC-WOZ corpus. The mechanistic claim is that depression is encoded in the unpredictability of how someone's voice changes across a conversation, not in its average properties. Both vocal dynamics papers converge on the same structural signal, which adds modest confidence despite their shared low reproducibility.

██████████ 0.9 depression-biomarkers Preprint

ReadSaveConnections

Across 28,800 synthetic patient profiles from four frontier LLMs, this study finds that models produce individually coherent-sounding patients but catastrophically misrepresent population-level distributions—variance compression ranges from 14% to 62%, effectively erasing the clinical extremes that matter most for treatment design. Worse, 36.7% of simulated cases cross diagnostic thresholds between repeated runs, even when overall correlations appear stable. This is a direct warning against using LLM-simulated cohorts for clinical trial design or training data generation.

██████████ 0.9 digital-therapeutics Preprint

ReadSaveConnections

This study identifies a reproducible monocyte activation program triggered by brain injury that correlates with downstream clinical outcomes, using transcriptomics and flow cytometry. The relevance to mental health is that the same neuroinflammatory immune pathways activated by brain injury are implicated in depression and treatment resistance—understanding how monocytes are reprogrammed after CNS insult could explain why brain injury dramatically elevates psychiatric risk. The conserved nature of the program across cohorts strengthens its candidacy as a translational biomarker target.

██████████ 0.8 neuroinflammation Peer-reviewed

Read

ADAPTS uses a mixture-of-LLM-agents architecture to automatically rate depression and anxiety severity from long clinical interview transcripts, decomposing the task symptom-by-symptom rather than asking a single model to produce a holistic score. On high-discrepancy interviews it actually outperforms original human raters (absolute error 22 vs. 26), and achieves ICC of 0.877 with an extended protocol. The practical implication is that automated rating of interviews could scale clinical research and provide consistent scoring, though the closed datasets and undisclosed model identities make independent validation impossible at present.

██████████ 0.8 depression-biomarkers Preprint

ReadSaveConnections

A meta-analysis across more than 100 studies finds that undefended choices about how to measure outcomes—which scale to use, which subscale, which cutoff—reduce the rate of statistically significant findings by approximately 50%. This is a field-wide methodological warning: much of the heterogeneity in mental health research replication failures may not be due to true effect variability but to measurement flexibility that is never justified or disclosed. The implication is that pre-registration of measurement decisions should be treated as essential, not optional.

██████████ 0.8 depression-biomarkers Peer-reviewed

Read

Two vision-language models (Phi-3.5-Vision and Qwen2-VL) were evaluated for depression detection using face, voice, and text together, revealing large performance gaps between controlled lab data (80.4% accuracy) and naturalistic clinical recordings (33.9%), plus systematic racial and gender biases that differ by model. The paper introduces a fairness prompting framework using chain-of-thought and counterfactual loss, but the gap between lab and real-world performance is the most important signal: deployment conditions shatter the accuracy numbers typically reported in benchmark papers.

██████████ 0.8 depression-biomarkers Preprint

ReadSaveConnections

This PhD thesis shows that task-based EEG during a stop-signal task outperforms resting-state EEG for classifying adult ADHD, with gamma-band power over fronto-central and parietal regions as the strongest predictors. A separate DTI analysis links white matter integrity in motor-planning tracts to computationally modeled effort and reward sensitivity parameters. The value here is methodological: it demonstrates that engaging the brain during a relevant task, then extracting theory-driven computational parameters, beats passive resting-state recording—a lesson applicable across psychiatric biomarker research.

██████████ 0.8 computational-psychiatry Preprint

ReadSaveConnections

This paper proposes projecting natural language responses onto semantic axes derived from validated clinical scales (CES-D, STAI-Y) using Sentence-BERT embeddings, generating continuous depression and anxiety scores without any supervised training. The approach shows strong correlations with PHQ-9 and GAD-7, and critically, response format matters: asking people to write words or phrases outperforms free-text for measurement quality. The unsupervised design means it could generalize across languages and populations without retraining, which is meaningful for low-resource clinical settings.

██████████ 0.7 depression-biomarkers Preprint

ReadSaveConnections